Figure 1 - available via license: Creative Commons Attribution-ShareAlike 4.0 International
Content may be subject to copyright.
Relative performance degradation at various pooling factors using 16-bit vectors with HNSW indexing.
Source publication
Over the last few years, multi-vector retrieval methods, spearheaded by ColBERT, have become an increasingly popular approach to Neural IR. By storing representations at the token level rather than at the document level, these methods have demonstrated very strong retrieval performance, especially in out-of-domain settings. However, the storage and...
Context in source publication
Context 1
... the interest of simplicity and space, we only report the results of sequential pooling for factors 2 and 4, as the observed performance degradation varies wildly between datasets and increases is too quickly for this approach to be viable in comparison to the other ones. Worth noting however that, despite its overall noticeably weaker performance, sequential pooling performs remarkably strongly on the scidocs dataset, reaching the strongest performance at a pooling factor of 2, before degrading behind the other two methods at factor 4. An overview of the relative performance degradation of the best-performing pooling method, hierarchical pooling, across pooling factors on all the small-sized evaluation datasets can be found in Figure 1 . ...
